{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:PBV5SDYOUULUA74I3UFVKCXSUE","short_pith_number":"pith:PBV5SDYO","canonical_record":{"source":{"id":"2605.13831","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T17:52:53Z","cross_cats_sorted":[],"title_canon_sha256":"5fa28a18467e969ced6bf97c59d63cb0c667de6bda223dd54c3d14045a5e5428","abstract_canon_sha256":"6f320a328b6b959c0c05b4b9acea7afc9b3a7b763f427d3c7f4cb1b1daadb8fa"},"schema_version":"1.0"},"canonical_sha256":"786bd90f0ea517407f88dd0b550af2a124ff19e0826b8690315019ac4872e692","source":{"kind":"arxiv","id":"2605.13831","version":1},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.13831","created_at":"2026-05-18T02:44:15Z"},{"alias_kind":"arxiv_version","alias_value":"2605.13831v1","created_at":"2026-05-18T02:44:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13831","created_at":"2026-05-18T02:44:15Z"},{"alias_kind":"pith_short_12","alias_value":"PBV5SDYOUULU","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"PBV5SDYOUULUA74I","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"PBV5SDYO","created_at":"2026-05-18T12:33:37Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:PBV5SDYOUULUA74I3UFVKCXSUE","target":"record","payload":{"canonical_record":{"source":{"id":"2605.13831","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T17:52:53Z","cross_cats_sorted":[],"title_canon_sha256":"5fa28a18467e969ced6bf97c59d63cb0c667de6bda223dd54c3d14045a5e5428","abstract_canon_sha256":"6f320a328b6b959c0c05b4b9acea7afc9b3a7b763f427d3c7f4cb1b1daadb8fa"},"schema_version":"1.0"},"canonical_sha256":"786bd90f0ea517407f88dd0b550af2a124ff19e0826b8690315019ac4872e692","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-18T02:44:15.062912Z","signature_b64":"kfrSOrhF/KD2l/4lEApRv9og2CkkCBZno1GXMlgSsGgrloKzgZ/i2XPJF9p7J5DWwJTDYWOJ/to2E0EMDT8UBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"786bd90f0ea517407f88dd0b550af2a124ff19e0826b8690315019ac4872e692","last_reissued_at":"2026-05-18T02:44:15.061085Z","signature_status":"signed_v1","first_computed_at":"2026-05-18T02:44:15.061085Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2605.13831","source_version":1,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T02:44:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"Mq+Og1q+oLEhqbp9evUAKNsDhivXv+7XlJ6fTh6xnRKN85LSPHDgiVYXZm2lvtTSVE5xG+/bdLZd9odSGuQDCA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-06T08:56:30.452563Z"},"content_sha256":"08fa508b72dc45dc0286a0b993f46234e092165e2490f58f0efff41e8366fb31","schema_version":"1.0","event_id":"sha256:08fa508b72dc45dc0286a0b993f46234e092165e2490f58f0efff41e8366fb31"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:PBV5SDYOUULUA74I3UFVKCXSUE","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Balanced long-document VQA data lets a vision-language model trained to 128K generalize to 512K contexts without further training.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Chaoyi Huang, Haodong Duan, Ji Luo, Lishu Luo, Shen Yan, Shuai Peng, Sihang Yuan, Sijin Wu, Weiwei Liu, Yangqiu Song, Yi Lin, Zhaowei Wang","submitted_at":"2026-05-13T17:52:53Z","abstract_excerpt":"Long-context modeling is becoming a core capability of modern large vision-language models (LVLMs), enabling sustained context management across long-document understanding, video analysis, and multi-turn tool use in agentic workflows. Yet practical training recipes remain insufficiently explored, particularly for designing and balancing long-context data mixtures. In this work, we present a systematic study of long-context continued pre-training for LVLMs, extending a 7B model from 32K to 128K context with extensive ablations on long-document data. We first show that long-document VQA is subs"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"MMProLong improves long-document VQA scores by 7.1% and maintains strong performance at 256K and 512K contexts beyond its 128K training window, without additional training.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the observed gains and generalization stem primarily from the specific long-document VQA mixture and balanced length distribution rather than from the base model choice, total token budget, or unablated factors.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Continued pre-training with balanced long-document VQA data extends a 7B LVLM to 128K context, improving long-document VQA by 7.1% and generalizing to 512K without further training.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Balanced long-document VQA data lets a vision-language model trained to 128K generalize to 512K contexts without further training.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"6f820b99385b1fe95564c3922b145fdc6135c7c26f9e6a59a51c7ea8687b3e9a"},"source":{"id":"2605.13831","kind":"arxiv","version":1},"verdict":{"id":"cb2a87b0-c504-4a58-a433-c73cde3a5a50","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:12:55.734913Z","strongest_claim":"MMProLong improves long-document VQA scores by 7.1% and maintains strong performance at 256K and 512K contexts beyond its 128K training window, without additional training.","one_line_summary":"Continued pre-training with balanced long-document VQA data extends a 7B LVLM to 128K context, improving long-document VQA by 7.1% and generalizing to 512K without further training.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the observed gains and generalization stem primarily from the specific long-document VQA mixture and balanced length distribution rather than from the base model choice, total token budget, or unablated factors.","pith_extraction_headline":"Balanced long-document VQA data lets a vision-language model trained to 128K generalize to 512K contexts without further training."},"references":{"count":80,"sample":[{"doi":"","year":2025,"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","ref_index":1,"cited_arxiv_id":"2505.09388","is_internal_anchor":true},{"doi":"","year":2025,"title":"The llama 4 herd: The beginning of a new era of natively multimodal ai innovation, 2025.URL https://ai.meta.com/blog/llama-4-multimodal-intelligence/, 2025","work_id":"11252c0b-f5cb-4a59-bf5f-303a4578a0dc","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Qwen3-VL Technical Report","work_id":"1fe243aa-e3c0-4da6-b391-4cbcfc88d5c0","ref_index":3,"cited_arxiv_id":"2511.21631","is_internal_anchor":true},{"doi":"","year":2025,"title":"Bytedance Seed. Seed2. 0 model card: Towards intelligence frontier for real-world complexity. Technical report, Technical report, Bytedance, 2025. URL https://lf3-static. bytednsdoc. com","work_id":"c557bd47-0b12-4507-8298-047a12925268","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Mmlong- bench: Benchmarking long-context vision-language models effectively and thoroughly","work_id":"7c57f3bc-9274-4cae-9408-0472fb16354a","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":80,"snapshot_sha256":"f787925976b8190d19cef20e2450c562ce76365a3b7eb80af780d5141619c584","internal_anchors":21},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"cb2a87b0-c504-4a58-a433-c73cde3a5a50"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-18T02:44:15Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"rp8mu0wpvENEBPCoDvZzy78r0j+waAmxyYCLO26NC9xqSTdZr3OAyToNmckaS6wLSHGnej/xl/3oveBvrUrOBQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-06-06T08:56:30.453171Z"},"content_sha256":"daef4b53c194e29c14e324851c125ab960607c4b96f971b1fb3bf777eb022b5d","schema_version":"1.0","event_id":"sha256:daef4b53c194e29c14e324851c125ab960607c4b96f971b1fb3bf777eb022b5d"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/PBV5SDYOUULUA74I3UFVKCXSUE/bundle.json","state_url":"https://pith.science/pith/PBV5SDYOUULUA74I3UFVKCXSUE/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/PBV5SDYOUULUA74I3UFVKCXSUE/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-06-06T08:56:30Z","links":{"resolver":"https://pith.science/pith/PBV5SDYOUULUA74I3UFVKCXSUE","bundle":"https://pith.science/pith/PBV5SDYOUULUA74I3UFVKCXSUE/bundle.json","state":"https://pith.science/pith/PBV5SDYOUULUA74I3UFVKCXSUE/state.json","well_known_bundle":"https://pith.science/.well-known/pith/PBV5SDYOUULUA74I3UFVKCXSUE/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:PBV5SDYOUULUA74I3UFVKCXSUE","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"6f320a328b6b959c0c05b4b9acea7afc9b3a7b763f427d3c7f4cb1b1daadb8fa","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T17:52:53Z","title_canon_sha256":"5fa28a18467e969ced6bf97c59d63cb0c667de6bda223dd54c3d14045a5e5428"},"schema_version":"1.0","source":{"id":"2605.13831","kind":"arxiv","version":1}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.13831","created_at":"2026-05-18T02:44:15Z"},{"alias_kind":"arxiv_version","alias_value":"2605.13831v1","created_at":"2026-05-18T02:44:15Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.13831","created_at":"2026-05-18T02:44:15Z"},{"alias_kind":"pith_short_12","alias_value":"PBV5SDYOUULU","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_16","alias_value":"PBV5SDYOUULUA74I","created_at":"2026-05-18T12:33:37Z"},{"alias_kind":"pith_short_8","alias_value":"PBV5SDYO","created_at":"2026-05-18T12:33:37Z"}],"graph_snapshots":[{"event_id":"sha256:daef4b53c194e29c14e324851c125ab960607c4b96f971b1fb3bf777eb022b5d","target":"graph","created_at":"2026-05-18T02:44:15Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"MMProLong improves long-document VQA scores by 7.1% and maintains strong performance at 256K and 512K contexts beyond its 128K training window, without additional training."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That the observed gains and generalization stem primarily from the specific long-document VQA mixture and balanced length distribution rather than from the base model choice, total token budget, or unablated factors."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"Continued pre-training with balanced long-document VQA data extends a 7B LVLM to 128K context, improving long-document VQA by 7.1% and generalizing to 512K without further training."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"Balanced long-document VQA data lets a vision-language model trained to 128K generalize to 512K contexts without further training."}],"snapshot_sha256":"6f820b99385b1fe95564c3922b145fdc6135c7c26f9e6a59a51c7ea8687b3e9a"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"paper":{"abstract_excerpt":"Long-context modeling is becoming a core capability of modern large vision-language models (LVLMs), enabling sustained context management across long-document understanding, video analysis, and multi-turn tool use in agentic workflows. Yet practical training recipes remain insufficiently explored, particularly for designing and balancing long-context data mixtures. In this work, we present a systematic study of long-context continued pre-training for LVLMs, extending a 7B model from 32K to 128K context with extensive ablations on long-document data. We first show that long-document VQA is subs","authors_text":"Chaoyi Huang, Haodong Duan, Ji Luo, Lishu Luo, Shen Yan, Shuai Peng, Sihang Yuan, Sijin Wu, Weiwei Liu, Yangqiu Song, Yi Lin, Zhaowei Wang","cross_cats":[],"headline":"Balanced long-document VQA data lets a vision-language model trained to 128K generalize to 512K contexts without further training.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T17:52:53Z","title":"Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context"},"references":{"count":80,"internal_anchors":21,"resolved_work":80,"sample":[{"cited_arxiv_id":"2505.09388","doi":"","is_internal_anchor":true,"ref_index":1,"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":2,"title":"The llama 4 herd: The beginning of a new era of natively multimodal ai innovation, 2025.URL https://ai.meta.com/blog/llama-4-multimodal-intelligence/, 2025","work_id":"11252c0b-f5cb-4a59-bf5f-303a4578a0dc","year":2025},{"cited_arxiv_id":"2511.21631","doi":"","is_internal_anchor":true,"ref_index":3,"title":"Qwen3-VL Technical Report","work_id":"1fe243aa-e3c0-4da6-b391-4cbcfc88d5c0","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":4,"title":"Bytedance Seed. Seed2. 0 model card: Towards intelligence frontier for real-world complexity. Technical report, Technical report, Bytedance, 2025. URL https://lf3-static. bytednsdoc. com","work_id":"c557bd47-0b12-4507-8298-047a12925268","year":2025},{"cited_arxiv_id":"","doi":"","is_internal_anchor":false,"ref_index":5,"title":"Mmlong- bench: Benchmarking long-context vision-language models effectively and thoroughly","work_id":"7c57f3bc-9274-4cae-9408-0472fb16354a","year":2025}],"snapshot_sha256":"f787925976b8190d19cef20e2450c562ce76365a3b7eb80af780d5141619c584"},"source":{"id":"2605.13831","kind":"arxiv","version":1},"verdict":{"created_at":"2026-05-14T19:12:55.734913Z","id":"cb2a87b0-c504-4a58-a433-c73cde3a5a50","model_set":{"reader":"grok-4.3"},"one_line_summary":"Continued pre-training with balanced long-document VQA data extends a 7B LVLM to 128K context, improving long-document VQA by 7.1% and generalizing to 512K without further training.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"Balanced long-document VQA data lets a vision-language model trained to 128K generalize to 512K contexts without further training.","strongest_claim":"MMProLong improves long-document VQA scores by 7.1% and maintains strong performance at 256K and 512K contexts beyond its 128K training window, without additional training.","weakest_assumption":"That the observed gains and generalization stem primarily from the specific long-document VQA mixture and balanced length distribution rather than from the base model choice, total token budget, or unablated factors."}},"verdict_id":"cb2a87b0-c504-4a58-a433-c73cde3a5a50"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:08fa508b72dc45dc0286a0b993f46234e092165e2490f58f0efff41e8366fb31","target":"record","created_at":"2026-05-18T02:44:15Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"6f320a328b6b959c0c05b4b9acea7afc9b3a7b763f427d3c7f4cb1b1daadb8fa","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T17:52:53Z","title_canon_sha256":"5fa28a18467e969ced6bf97c59d63cb0c667de6bda223dd54c3d14045a5e5428"},"schema_version":"1.0","source":{"id":"2605.13831","kind":"arxiv","version":1}},"canonical_sha256":"786bd90f0ea517407f88dd0b550af2a124ff19e0826b8690315019ac4872e692","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"786bd90f0ea517407f88dd0b550af2a124ff19e0826b8690315019ac4872e692","first_computed_at":"2026-05-18T02:44:15.061085Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-18T02:44:15.061085Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"kfrSOrhF/KD2l/4lEApRv9og2CkkCBZno1GXMlgSsGgrloKzgZ/i2XPJF9p7J5DWwJTDYWOJ/to2E0EMDT8UBg==","signature_status":"signed_v1","signed_at":"2026-05-18T02:44:15.062912Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.13831","source_kind":"arxiv","source_version":1}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:08fa508b72dc45dc0286a0b993f46234e092165e2490f58f0efff41e8366fb31","sha256:daef4b53c194e29c14e324851c125ab960607c4b96f971b1fb3bf777eb022b5d"],"state_sha256":"71f89829dadda1e928cca8d5efc53f53a45ec1d0217ed59ccad4ae6eef879175"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"0IHkjrn9gXuqM6yaLpfsEhlT8WzFGj2IcD+svA4I+2kM6VV1YXNFxFqxrvU8zG5Hrnvk3VY5DvleS7XSAvVbAQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-06-06T08:56:30.457008Z","bundle_sha256":"14b99455083672e84cddbd48b05c9d5f93070d548a5b43166adce323dba58fbd"}}